A three component learned model had higher mean relative error than a cosine model for 50 to 250 ms SEEG prediction in five patients

We tested whether response shapes from other stimulation shafts improve prediction from five pulses. The public SEEG data had five patients and 1,081 eligible links. The learned model and fixed cosine model each used three components. The time window was 50 to 250 ms. The learned model had 2.54 percent higher error on the equal patient relative score. It improved the score in two patients. The exact patient sign flip test gave p = 0.625. The empirical bootstrap limits for error reduction were -15.40 to 8.66 percent. These limits describe five observed patients. The test requires independent patients and symmetric effects under the zero null. The package includes input windows, code, fitted bases, all stored split scores, patient results, and a raw data retrieval recipe. A fresh extraction from 120 hashed public MEF files reproduced all 3,477,760 waveform values. A separate calculation checked all 3,813,768 stored error values. A repeat analysis reproduced every table and array. This is a method comparison within the same five patients used in earlier work. It does not establish a population effect.

stereoelectroencephalography · electrical stimulation · response prediction · principal components · cosine basis · cross-validation · prediction error